面向属性情感分类的无标签数据增强位置感知网络  

Unlabeled Data Enhanced Position-Aware Network for Aspect-based Sentiment Classification

作　　者：蒋慧凯 李晓戈 安晓春 王甜甜[3] 阮桁 JIANG Huikai;LI Xiaoge;AN Xiaochun;WANG Tiantian;RUAN Heng(School of Computer Science and Technology,Xi'an University of Posts and Telecommunications,Xi'an 710121,China;Shaanxi Key Laboratory of Network Data Analysis and Intelligent Processing,Xi'an 710121,China;School of Management,Shandong University,Jinan 250100,China)

机构地区：[1]西安邮电大学计算机学院,西安710121 [2]陕西省网络数据分析与智能处理重点实验室,西安710121 [3]山东大学管理学院,济南250100

出　　处：《计算机工程》2023年第11期106-114,共9页Computer Engineering

基　　金：陕西省重点研发计划(2020ZDLGY09-05)。

摘　　要：在现有的属性情感分类研究中,训练模型时大多完全依赖标签数据或需要引入文本级标签数据作为补充,很少关注无标签数据对模型性能的提升。提出一种基于无标签数据增强的位置感知网络(UDE-PAN)。引入交叉可视训练(CVT)的半监督训练算法,使模型能够同时利用无标签数据。CVT算法在标签数据和无标签数据中交替训练模型来提升表征学习能力,使模型在无标签数据中学习到额外知识。此外,基于语义相对距离(SRD)嵌入层和动态特征加权(CDW)层捕获属性词和上下文的关系:SRD嵌入层显式地将位置信息建模成特征向量,使上下文特征包含特定的属性信息;CDW层通过动态设置权重系数来感知上下文中与属性词更密切的部分。在SemEval14的2个公开数据集上的实验结果表明:UDE-PAN的准确率分别达到76.23%、82.47%,Macro-F1值分别达到72.13%、73.97%,优于对比的主流模型,验证了模型的有效性;借助CVT算法,无标签数据的训练对模型的准确率平均提升1%,Macro-F1平均提升2%,验证了无标签数据可以有效增强模型性能。Recently,in Aspect-based Sentiment Classification(ASC)studies,models were trained with labeled data or document-level data as a supplement;however,little attention was given to exploring the improvements in model performance using unlabeled data.Accordingly,the Unlabeled Data Enhanced Position-Aware Network(UDE-PAN)is proposed herein to improve the ASC performance.Specifically,a semi-supervised training algorithm,Cross-View Training(CVT)algorithm,is introduced that enables the proposed model to utilize unlabeled data.The CVT algorithm alternatively trains the model on labeled and unlabeled data,improving the model's ability to learn sentence representations.Moreover,Semantic-Relative Distance(SRD)embedding and Context features Dynamic Weighting(CDW)layers are adopted to learn the relationship between aspect words and context.The SRD embedding layer explicitly models the relative position information,so the context features contain more specific information.The CDW layer captures the part closer to the aspect in the context through the dynamic weighting of the context features.Finally,extensive experiments are conducted considering two public benchmark datasets from SemEval14.As a result,the accuracies of UDE-PAN of 76.23%and 82.47%are achieved and the Macro-F1 are 72.13%and 73.97%,respectively.This demonstrates better performance than the comparison models and proves the effectiveness of the proposed model.With the help of CVT algorithm,the training of unlabeled data improves the accuracy of the model by the average of 1%and Macro-F1 by the average of 2%,verifying that unlabeled data can effectively enhance the performance of the model.

关 键 词：属性情感分类 无标签数据 位置感知 交叉视图训练 注意力机制 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]